On sampling generating sets of finite groups Sergey Bratus∗, Igor Paky Abstract. Let G be a finite group. For a given k, what is the probability that a group is generated by k of its random elements? How can one uniformly sample these generating k-tuples? In this paper we answer these questions for the class of nilpotent groups. Applications to product replacement algorithms and random random walks are discussed. 1. Introduction Let G be a finite group. A sequence of group elements (g1; : : : ; gk) is called a generating k-tuple of G if these elements generate G (we write hg1; : : : ; gki = G). Let Nk(G) be the set of all generating k-tuples of G, and let Nk(G) = jNk(G)j. We consider two related problems on generating k-tuples. Given G and k > 0, 1) Determine Nk(G). 2) Generate a random element of Nk(G), so that each element of Nk(G) appears with probability 1=Nk(G). The problem of determining the structure of Nk(G) is of interest in several contexts. The counting problem goes back to Hall, who expressed Nk(G) as a M¨obiustype summation of Nk(H), taken over all maximal subgroups H ⊂ G (see [33]). Recently, the counting problem has been studied for large simple groups, where remarkable progress has been made (see [39, 40, 41, 49]). In this paper we analyze Nk(G) for nilpotent groups. We also show that Nk(G) r minimizes when G ' Z2, r ≥ log2 jGj. The sampling problem, while often used in theory as a tool for approximate counting, recently began a life of its own. In [14] Celler et al. proposed a product replacement Markov chain on Nk(G), which is believed to be rapidly mixing. The subject was further investigated in [8, 15, 21, 22, ?]. We present an efficient and economical algorithm for sampling in case when G is nilpotent. ∗Department of Mathematics, Northeastern University, Boston, MA 02115, E-mail: [email protected] yDepartment of Mathematics, Yale University, New Haven, CT 06520, E-mail: [email protected] 2 The generating k-tuples also occur in connection with the so-called random random walks, which are ordinary random walks on G with random generating sets. The analysis of these \average case" random walks was inspired by Aldous and Diaconis in [1] and was continued in a number of papers (see e.g. [26, 52, 47, 55]). We will show that one can use the sampling problem to simulate these random random walks. 2. Definitions and main results 2.1. Counting problem Let G be a finite group. By jGj denote the order of G. As in the introduction, let Nk(G) = jNk(G)j be the number of generating k-tuples hg1; : : : ; gki = G. It is often convenient to consider the probability 'k(G) that k uniform independent group elements generate G : N (G) ' (G) = k k jGjk Theorem 1. For any finite group G, 1 > > 0, we have 'k(G) > 1 − given k > log2 jGj + maxf3; 2 log2 1/g. This is a slight improvement over a more general classical result by Erd}os and R´enyi in [28] (see also [27]). Define {(G) to be the smallest possible number of generators of G. In other words, let {(G) = minfk j Nk(G) > 0g: The problem of evaluating {(G) has been of intense interest for classes of groups as well as for individual groups (see [17]). It is known that {(G) = 2 for all simple, nonabelian groups, and that n=2 {(G) ≤ n=2 for G ⊂ Sn, with equality achieved when G ' Z2 , and n is n even. Also, it is easy to see that {(G) ≤ log2 jGj, with equality for G ' Z2 . Let #(G) be the smallest k such that at least 1=3 of the random k-tuples (g1; : : : ; gk) generate the whole group. In other words, let 1 #(G) = min k j ' (G) > : k 3 Note that Theorem 1 immediately implies that #(G) ≤ log2 jGj + 3 By definition #(G)={(G) > 1. It is unclear, however, how big this ratio can be (see [49, ?]). 3 Here are few known results. When G is simple, it is known that '2(G) ! 1 as jGj ! 1 (see [53]). This was the famous Kantor{Lubotzky conjecture, now a theorem. For G = An, this is a famous result of Dixon (see [25]). For classical simple groups of Lie type the result was confirmed by Kantor and Lubotzky (see [39]). In full generality it was recently proved by Liebeck and Shalev (see [40]). This immediately implies that #(G) < C for any simple group G and some universal constant C. It was noted in [22] that when G is nilpotent, then '{+1(G) > Const. The following result is an improvement. Theorem 2. Let G be a finite nilpotent group. Then #(G) ≤ {(G) + 1. We refer the reader to [23, 49, ?] for further discussion. 2.2. On presentations of groups There are several ways a finite group G can be presented as input to an algorithm. Any group{theoretic algorithm needs to be able to perform the group operation, to find inverse elements and to compare the results of these operations with the identity element of G. The complexity of an algorithm is expressed as a function of the times necessary to perform these operations and other parameters. Thus, regardless of the presentation of G, denote by µ the time necessary for group operations (multiplication, taking an inverse, comparison with id1). Further, randomized algorithms usually assume the ability to generate random elements of the group G. Denote by ρ the complexity of generating a (nearly) uniform group element (call this the random generation subroutine). It is also convenient to denote by η the time required to check whether given k group elements generate the entire group. We call this task a generation test. We start with permutation groups, which are defined as subgroups of a per- mutation group Sn. The group is presented by a set of generators. This is the best understood class of groups with efficient management, random elements generation, generation test, etc., based on the fundamental algorithms by C. Sims (see e.g. [54, 16, 43]. In particular one has ρ = O(µn), and η = O(µn4) (one can show that in this case by reducing the problem to group membership). A matrix group is a group defined as a subgroup of GL(n; q). This is a harder class of groups to work with (see [37, 8]). Recently some important advances have been made in this setting (see [10, 13, 45, 42]). Still, polynomial time management for matrix groups is yet to be discovered. One of the most general and widely accepted is the black box setting (see [8]), in which group elements are encoded by bit strings of a uniform fixed length n (possibly non-uniquely, i.e. several different strings may correspond to the same element of G). A black box oracle is given, that can multiply elements, take their inverses (returning the results as bit strings of the same 1For some presentations, such as the presentation by generators and relations, the latter task can be non-trivial. The black box model discussed below makes the assumption that the identity test, i.e. comparison with id, can be performed efficiently. 4 encoding), and compare elements with identity (see [8]), in time polynomial in n. Note that n gives a bound on log2 jGj. This presentation of a group generalizes both permutation groups and matrix groups. This setting proved itself to be useful for working with general finite groups about which we have limited information. In his pioneering work [6], Babai was able to find a polynomial time algo- rithm for generating (nearly) uniform group elements. The product replace- ment algorithm of [14] was designed to give a practical algorithm for random generation. These algorithms were used in a number of subsequent works, par- ticularly on recognition of various classes of finite groups (see [11, 12, 38, 45]). Following Babai (see [6]), there is no subexponential in n algorithm which can perform the general generation test. When necesary, we isolate the complexity of performing the generation test in a separate subroutine of complexity η. Finally, a finite solvable group G can be given by a polycyclic generating sequence, also referred to as an AG-system (see [54], Sec. 9.4). In this case, both the random generation subroutine and the generation test can be easily performed in nearly linear time. While no subexponential algorithm for finding such a presentation is known, the existing algorithms implemented in GAP and MAGMA are known to be very efficient in practice. We will consider nilpotent groups that come in such a presentation. 2.3. Sampling problem Now consider the sampling problem (see introduc- tion) from the computational point of view. We immediately obtain the fol- lowing result. Proposition 2.1. Let G be a black box group with a generation test oracle, and a random generation oracle. Let ρ be the time required for the random generation oracle to generate a (nearly) uniform random element of G, and η be the time in which the generation test oracle can perform a generation test. Let k ≥ #(G). Then there exists a randomized algorithm for sampling from Nk(G) in time O(ρk + η). Indeed, given k ≥ #(G), we can always sample from Nk(G) by simply generating a uniform k-tuple and testing whether it generates the whole group G. We call this method Choose{and{Check. The sampling problem is open for {(G) ≤ k < #(G).